The line where every customer-facing AI breaks — and where Koni already stands
The whole industry is betting on the same thing, and it is precise about why it is hard. The figures below are from Dynatrace, The Pulse of Agentic AI 2026 (Research Report, ©2026 Dynatrace; survey of 919 enterprise leaders, ±3.2% margin of error):
- Per Dynatrace, customer-facing agents are the fastest-growing use case of the next five years (+31 points) — the niche everyone is racing toward.
- The report finds 69% of agentic decisions are still verified by a human, and 64% of teams build agents that are both autonomous and human-supervised. Only 13% trust full autonomy.
- Dynatrace names the #1 technical barrier to production almost verbatim: "limited ability to set clear rules for when an agent should act autonomously versus require human approval."
- The pilot-to-production gate, per the same study, is security (59%) and reliability/accuracy (55%) — anchoring a probabilistic model to deterministic, verifiable facts.
Source: Dynatrace, The Pulse of Agentic AI 2026, ©2026 Dynatrace. Survey of 919 enterprise leaders (±3.2%).
Read that as a scorecard for the big leagues. It is the exact place almost every agent stalls: it either over-acts and can't be trusted, or it can't tell when to stop.
Koni is already standing on that line — not in a sandbox, but inside a live phone call you can place right now.
A visitor calls the building. Koni answers in a natural voice, understands who they are there to see, resolves the exact home, and calls the resident. Then she does the hardest thing an agent can do: she stops. She stops at the precise boundary the industry calls the wall — the access decision — and hands it to the human who owns it: the resident. Koni knows where to stop (before the door), why she stops (no one but you decides who enters your home), and exactly how to continue (the instant the resident says yes, she bridges them; if no one answers, she says so and closes without ever guessing).
Autonomous up to the human. Human-in-the-loop for the decision. Deterministic on both sides. That is supervised autonomy, running live on a real carrier line — the exact architecture 64% of the market is still trying to build. This is phase one of Koni.
The black box behind the voice
A chatbot is allowed to guess. An agent that routes physical access is not. The voice is the half you see; the half that took the months is making the action deterministic — same input, same verifiable result, every single time — while keeping the voice warm and human, never robotic. Tuning Gemini hot enough to sound like a person yet controlled enough to never once improvise the routing — that razor is the whole engineering problem, and it is where the nights went:
- A deterministic resolver that turns a spoken name into one exact extension — and the hard-won discovery that our carrier (GoTo) and our own database are two separate sources of truth that do not sync on their own. A name changed in one is silence in the other. Every mismatch was a call that died with no explanation — until we made the reconciliation explicit and verified it against the live database on every call.
- A call-coordination engine that holds the visitor's line, rings the resident, survives no-answer, and bridges cleanly — deployed under a rule earned by breaking production: never touch a live call-router without a rollback image and a real start-up check first.
- Real telephony, not a mock: SIP, RTP, and live media that must actually connect across a real carrier, where one wrong parameter is twenty seconds of dead air and a lost visitor.
This is the boring, merciless layer under the intelligence — and it is the layer that decides whether an AI can be trusted to act in the physical world at all.
The triumph: we killed the call that dies with no trace
Every intercom on earth has the same disease. A visitor arrives, no one answers upstairs, the call falls into a carrier voicemail nobody checks — and it vanishes. No record anyone came. No accountability. We cured it.
Koni rings her own native app first — an endpoint we control — so nothing is swallowed by a mailbox no one reads. If the app doesn't answer, the call falls through to the resident's phone with a record attached. If still no one answers, Koni logs the arrival and takes the visitor's message. Every visit now leaves a mark — for the first time.
That is the real work of this build: AI + software + telephony + a native app fused into a single working tool that did not exist before — standing on real infrastructure (GoTo PBX, and our native iOS app, currently in App Store review), not on a slide.
The hardest lesson
Mid-build, our Gemini key was compromised and abused: 36,750 fraudulent requests in a single day, ~MXN $24,772 in charges. We rebuilt on a hardened project, capped quotas at the API level, and deployed an automated billing kill-switch — a Cloud Function wired to a budget that automatically disconnects billing if spend ever crosses the cap. It is live and evaluates real spend continuously (verifiable in our repository's evidence); it has not had to fire, because on the new hardened project spend has stayed well under the cap. Security for an agent that touches the real world is never finished — it is layered, and you pay for that lesson in real money and lost sleep.
Real traction: a client who pays today
Koni is not a prototype looking for its first user. She operates in production at the Los Helechos condominium as a billed service — invoiced at $3,500 MXN/month as its own "Koni intelligent-concierge service" line item, on SAT-certified CFDI invoices, with the July 2026 payment already cleared. This market is real, and it already pays. (Serving over 60,000 condominiums nationwide is the opportunity — Mexico alone has 35,219,141 dwellings, per INEGI's 2020 Census.)
Where Koni is going — the vision this foundation unlocks (not shipped here)
The voice flow is the ground floor of something far larger: an AI agent of absolute governance for the condominiums where millions of families co-own their buildings but have no one who knows every rule, forgets nothing, and hides nothing. The same deterministic discipline that keeps her from hallucinating on a phone call is what will let her be trusted with governance itself.
None of the following ships in this submission — it is the roadmap, stated as vision: a 2D animated face on the call flows; a new voice-native governance layer that lets Koni answer residents' rule questions by voice during a call; IoT water-leak detection with automatic shut-off valves; layered access security; and automated fees and payments.
We know exactly what this market needs, and we know how to fuse AI, software, hardware, PBX, and IoT into a tool of work that never existed. Phase one is live. You can call it right now — and it stops exactly where a human must decide. That is the future of customer-facing AI, and it is already answering the phone.
Built with
Gemini · Google Cloud · Node.js · Express · SIP / RTP telephony · GoTo PBX · native iOS app (in App Store review) · Azure · WebRTC
Log in or sign up for Devpost to join the conversation.